transformer-circuits

LIVE PULSE

3.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.1 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.7 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src3.9 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.1 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.7 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src

topic2 events

papersSEP 12 13:56 UTC

Anthropic Paper Proposes Mathematical Framework for Analyzing Transformer Circuits

Anthropic researchers published a paper outlining a mathematical approach to reverse-engineering how transformer models compute internally, treating attention heads and MLP layers as composable circuits. The framework aims to make the internal mechanisms of these models more tractable to study and explain. It is intended as a foundation for interpretability work rather than a description of any specific deployed system.

interpretability/transformers/research/anthropic

papersSEP 10 04:00 UTC

Training trajectories determine circuit removability in annealable soft-prior Transformers

Researchers asked whether retrieval circuits that small Transformers learn with the help of soft positional priors keep functioning once that prior is taken away. They tested this using a model whose prior-based attention biases can be gradually annealed out during training. The results indicate that the specific training trajectory, not just the architecture, decides whether a learned circuit can stand on its own after the prior is removed.

interpretability/transformers/training-dynamics/retrieval